Research On Lane Detection And Vehicle Detection Based On Monocular Vision

A set of Motor Vehicle Auto Driving System is of advancement,which can sense the whole body of a driving car.And it can not only fuse information,analyze information,carry out driverless operation finally,but also liberate the driver's hands,improve efficiency,and reduce traffic accidents.Therefore,it has received extensive attention at home and abroad.Although the traditional Auto Driving System uses many sensors such as radar,laser,and ultrasonic for the information around the car,it is difficult to be popular because of its high cost,single function,and inconformity with human perception.In recent years,with the development of computer vision technology,Auto Driving System based on monocular vision has been gradually recognized by people for its low cost,rich information,and intuitive results.Therefore,the detection of lane lines and vehicles is one of the basic tasks for it.So,it is of great research value and practical significance to carry out research on lane line and vehicle detection based on monocular vision.The main work of the paper is as follows:A lane detection algorithm based on cross-sectional features was proposed.According to rectangular form of the line in binary image,the points set were extracted.And the Connected Component Analysis was used to cluster,block and filter the points.The lane line was fitted by the straight-line and polynomial models respectively.And the Kalman filter was used to smooth detection results.Compared with the traditional lane detection method,experiments showed that this method could not only directly detect complex lines such as solid line,dotted line,straight line,and curve,but with high detection rate,good robustness,and good real-time performance.Vehicle detection training and detection algorithm based on Adaboost and Haar were improved.In the training stage,negative samples were divided into simple negative samples and complex negative samples.The first half of the Adaboost classifier was trained by positive samples and simple negative samples,followed by classifiers trained by using of positive samples and complex negative samples.Compared with the traditional training algorithm,this method could avoid premature divergence,as the structure of the classifier was simple and complex.Without affecting the efficiency of detection,the algorithm reduced the number of Haar features,and increased efficiency by 30%.In the detecting stage,the inverse perspective distance measurement algorithm was first used to calibrate.And the detection range of a candidate box of a certain row in the image was adaptively adjusted according to the actual vehicle width range and the longitudinal distance from the host vehicle.Compared with the traditional sliding window detection algorithm,this method filtered out more than half of the invalid detection windows and increased the detection speed by more than 7 times.A flash suppression algorithm was proposed.In the continuous video image detection,the result of the compressive tracking algorithm was merged with the corresponding target of the classifier.Two accumulators were set to count the number of occurrence and consecutive disappearance of the target.The results would be displayed only when the two accumulators met the requirements.And this result would be displayed as the target.Experiments showed that the algorithm could effectively distinguish four phenomena.And It was of good robustness,with both the stability of the detection window and the detection rate improved.